Islam (1995) - Growth Empirics A Panel Data Approach

Growth Empirics: A Panel Data Approach
Author(s): Nazrul Islam

Source: The Quarterly Journal of Economics, Vol. 110, No. 4 (Nov., 1995), pp. 1127-1170
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/2946651 .
Accessed: 25/04/2014 17:43
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to The Quarterly
Journal of Economics.
http://www.jstor.org
This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

All use subject to JSTOR Terms and Conditions
GROWTHEMPIRICS:A PANEL DATA APPROACH*

NAZRUL ISLAM
A panel data approach is advocated and implemented for studying growth
convergence. The familiar equation for testing convergence is reformulated as a
dynamic panel data model, and different panel data estimators are used to estimate
it. The main usefulness of the panel approach lies in its ability to allow for
differences in the aggregate production function across economies. This leads to
results that are significantly different from those obtained from single crosscountry regressions. In the process of identifying the individual "country effect," we
can also see the point where neoclassical growth empirics meets development
economics.
I. INTRODUCTION
In recent years there has been considerable empirical work on

cross-country growth. A close two-way relationship has been
observed between this work and the corresponding developments
in the theory of growth, in particular, the emergence of new
(endogenous) growth theories and the ensuing conflict between
these on the one hand and the preexisting models of growth in the
tradition of Solow [1956], Cass [1965], and Koopmans [1965], on
the other. A central focus of this work has been the issue of
convergence. While the finding of convergence has been generally
thought of as evidence in support of the Solow-Cass-Koopmans
model, absence of convergence has been regarded as supportive of
endogenous growth theories. The controversy has given rise to the
concept of "conditional convergence" meaning convergence after
differences in the steady states across countries have been controlled for.
A common feature of existing empirical studies on this issue
has been the assumption of identical aggregate production functions for all the countries. Although it has been correctly felt that
the production function may actually differ across countries,
efforts at allowing for such differences have been limited by the fact
that most of these studies have been conducted in the framework of
single cross-country regressions. In this framework it is econometri-
*I would like to thank Dale W. Jorgenson, Gary Chamberlain, Guido W.

Imbens, Robert M. Solow, James H. Stock, Lawrence F. Katz, Olivier J. Blanchard,
and an anonymous referee for their helpful suggestions. I also benefited from the
comments of Xoce Garcia, Eduard Sprokholt, and participants of the Harvard
Econometrics Lunch. All remaining errors are mine.
? 1995 by the President and Fellows of HarvardCollegeand the MassachusettsInstitute of
Technology.
The QuarterlyJournal of Economics,November1995

1128
QUARTERLYJOURNAL OF ECONOMICS
cally difficult to allow for such differences in the production

function as are not (easily) measurable.
The present paper advocates and implements a panel data
approachto deal with this issue. The panel data frameworkmakes
it possible to allow for differences of the above-mentionedtype in
the form of unobservableindividual "country effects." This paper
takes the recent work by Mankiw, Romer, and Weil [1992] as its
starting point and examines how the results change with the
adoption of the panel data approach. We reformulate the regression equation used in the study of convergence into a dynamic
panel data model with individual (country) effects and use the
panel data procedures to estimate it. This yields results that are
different from the corresponding results obtained from single
cross-sectionmethodology.First, the estimated rates of conditional
convergenceproveto be higher. Second, the estimated values of the
elasticity of output with respect to capital are found to be much
lower and more in conformity with its commonly accepted empirical values.
Investigation into the statistical sources of the aforementioned changes in the results shows that both of them can, to a
great extent, be explained in the framework of omitted variable
bias. The country-specific aspect of the aggregate production
function that is ignored in single cross-section regression, is
correlatedwith the includedexplanatoryvariables,and this creates
omitted variablebias. The panel data frameworkmakes it possible
to correct this bias. From growth theory's point of view, the panel
approachallows us to isolate the effect of "capital deepening" on
the one hand and technologicaland institutional differenceson the
other, in the process of convergence. The results indicate that
persistent differences in technology level and institutions are a
significantfactorin understandingcross-countryeconomicgrowth.
It becomes clear that if there had been no such differences, and
countries differed only in terms of capital per capita, convergence
would have proceededat a faster rate.
Contrary to what may appear at first sight, the finding of a
higher rate of conditionalconvergenceactually calls for more policy
activism. In the setup of identical productionfunctions, in order to
increase the steady state level of per capita income, countries were
to focus only on the rates of saving and labor force growth. But
when differencesin the aggregate productionfunction are allowed,
they are called upon to focus attention on all the tangible and
intangible factors that may enter into their respective individual

GROWTHEMPIRICS:A PANEL DATAAPPROACH
1129
(country) effects. Improvements in these factors may have direct

positive effects on the country's long-run income level. Our analysis shows that improvements in the country effect also lead to a
higher transitional growth rate. Furthermore, such improvements
may have a conducive effect on the traditional determinants of the
steady state level of income, i.e., saving rate and population growth
rate. Much of the discussion of development economics may be
thought to have been directed at ways to improve the countryspecific aspect of the aggregate production function. Explicit
recognition of this aspect by adopting the panel data framework,
therefore, creates a bridge between development economics and the
neoclassical empirics of growth.
The paper is organized as follows. In Section II we provide
further background on the issue of convergence. In Section III we
reformulate the growth equation as a dynamic panel data model. In
Section IV we discuss the relevant issues of panel estimation, and
the data and samples. Estimation results are presented in Section
V, and Section VI contains their interpretation. The analysis is
extended to include human capital in Section VII. In Section VIII
we present some analysis of the estimated country effects. Section
IX concludes.
II. THE ISSUE OF "CONVERGENCE"AND ITS EMPIRICALSEARCH
A major focus of recent work on growth empirics has been the

issue of convergence. The basic paradigm for this discussion had
been provided by the Solow [1956] model. The crucial assumption
in the Solow model of diminishing marginal returns to capital leads
the growth process within an economy to eventually reach the
steady state where per capita output, capital stock, and consumption grow at a common constant rate equaling the exogenously
given rate of technological progress. This led to the notion of
convergence, which in turn can be understood in two different
ways. The first is in terms of level of income. If countries are
similar in terms of preferences and technology, then the steady
state income levels for them will be the same, and with time they
will all tend to reach that level of per capita income. The second is
convergence in terms of the growth rate. Since in the Solow model
the steady state growth rate is determined by the exogenous rate of
the technological progress, then provided that technology is a
public good to be equally shared, all countries will eventually attain
the same steady state growth rate. The Cass-Koopmans version of

1130
the model, where the saving rate is dynamically optimized, also has
these implications.
It has now been quite some time that researchers have been
confronting real data with these hypotheses. Initially, much of this
work was conducted on the basis of the data of the developed
industrialized countries. Data availability had a significant role in
this choice of sample. In one of the recent works on this topic,
Baumol [1986], for example, reported finding convergence among a
group of countries included in Maddison's [1982] sample. These
countries tended to converge both to similar levels of per capita
income and to similar rates of growth.
An important question in this regard is what should be the
appropriate methodology for testing convergence. Since the notion
of convergence pertains to the steady states of the economies, a test
for convergence would require the assumption that the countries
included in the sample are in their steady states. However, judging
whether countries are in their steady states or not can be problematic. One way around this problem, therefore, is to study the
correlation between initial levels of income and subsequent growth
rates. Because of diminishing marginal returns to capital, countries with low levels of capital stock will have higher marginal
product of capital and hence, for similar saving rates, grow faster
than those with already higher levels of per capita capital stock.
Thus, a finding of negative correlation between initial levels of
income and subsequent growth rates has become a popular criterion for judging whether or not convergence holds. It may be noted
that this negative correlation has the scope of being interpreted as
evidence of convergence in terms of both income level and growth
rate. Poorer countries "catch up" (convergence in terms of income
level) with the richer countries by initially growing faster, and then
their growth rates slow down to the common rate of technological
progress (resulting in convergence in growth rates).
As more wide-ranging data sets became available, empirical
regularities of the growth process over a wider cross section of
countries started to draw the attention of researchers. Romer
[1989a] has been influential in drawing the attention of macroeconomists to the fact that over a large sample of countries, the
correlation between initial income levels and subsequent growth
rates is either zero or even positive. The evidence has also been
interpreted as one of "persistence" of significant differences in
income level and growth rates among countries [Rebelo 1991; King

1131
and Rebelo 1989]. The rise of endogenous growth theories, as is

known, has been, to a large extent, a response to these empirical
findings.
A different response to these same facts has been the proposition of the concept of "conditionalconvergence."Barro,in his first
empiricalwork [1989] on growth, showed that if differencesin the
initial level of human capital (along with some other pertinent
variables) are controlled for, then the correlation between the
initial level of income and subsequent growth rate turn out to be
negative even in the wider sample of countries. This concept of
conditional convergence found its more explicit formulation in
Barro and Sala-i-Martin [1992] and Mankiw, Romer, and Weil
[1992]. Both these papers emphasizedthe fact that the neoclassical
growth model (either Solow's or its optimal saving version by Cass
and Koopmans) did not imply that all countries would reach the
same level of per capita income. Instead, what it implied is that
countries would reach their respective steady states. Hence, in
looking for convergencein a cross-countrystudy, it is necessary to
control for the differencesin steady states of differentcountries.
In the Solow version of the neoclassicalmodel, the steady state
income level of a country is determinedby the country's saving and
labor force growth rates (which are treated as exogenous), and
some other parameters of technology (including the depreciation
rate). For the Cass-Koopmansversion of the model, the steady
state is determined by the underlying parameters describing the
preference and technology of the country. Both Barro and Sala-iMartin (hereinafter B-S) and Mankiw, Romer, and Weil (hereinafter M-R-W)found strong evidence for conditional convergence. In
M-R-W, which proceeded from the original Solow model, differences in the steady state income levels across countries were
controlledfor by the inclusion of saving and populationgrowth rate
variables in the regression. B-S, on the other hand, worked with
the optimal saving version of the neoclassical model, and hence
they had to consider such measurable variables as could proxy for
the underlyingparametersof preferenceand technology. However,
since the main aim of B-S was to study convergence among the
United States, they could assume that preference and technology
were uniform across the states resulting in the same steady state.
Their regressions for the states, therefore, did not include variables
designed to control for differencesin steady states (except for some
regional dummies).

1132
The tests for convergence conducted so far have followed a

similar methodology. They basically consist of running crosssection regressions with the subsequent growth rate as the dependent variable and the initial level of income as the prime explanatory variable. Other variables appearingon the right-hand side of
the regressions are designed to control for the differences in
preference and technology and hence in steady states. One difficulty with this methodology is that only such differences in
preference and technology can be accounted for as can be properly
observed and measured. Yet differencesin preferenceand technology across countries have dimensions that are not readily measurable or observable.In the frameworkof cross-section regression, it
is not possible to take account of such unobservableor unmeasurable factors. Only a panel data approach can overcome this
problem.
III. GROWTHREGRESSIONAS A DYNAMICPANEL DATA MODEL
The usefulness of a panel data approachcan be illustrated on

the basis of the work by M-R-W.They started with the following
"textbook Solow model" featuring the Cobb-Douglasproduction
function with labor-augmentingtechnologicalprogress:
(1)
Y(t) = K(t)0(A(t)L(t))1-a
0 < a < 1,
where Y is output, K is capital, and L is labor. L and A are assumed
to grow exogenously at rates n and g so that

L(t) = L(O)ent
A(t) = A(0)egt.
Assuming that s is the constant fraction of output that is saved and

invested, and defining output and stock of capital per unit of
effective labor as 9 = YIAL and k = K/AL, respectively, the
dynamicequation for k is given by
(2)
k(t) = s9(t) - (n + g + 6)k(t)

= sk(t)a - (n + g + 6)k(t),
where 8 is the constant rate of depreciation. It is evident that k

convergesto its steady state value:
k*=

1133
Upon substitution this gives the following expression for steady

state per capita income:
(3)
Y(t)1
in =
lnA(O) + gt +
In (s) +-
In (n
g +
Assuming that the countries are currently in their steady states,

M-R-Wused this equation to see how differing saving and labor
force growth rates can explain the differences in the current per
capita incomes across countries. In general, they found the model
to be quite successful in explaining a large fraction of the crosscountry variations in income, but the estimates of the elasticity of
output with respect to capital, a, were found to be unusually high.
One approachto explaining this type of result (high a) has been to
argue that capital in the productionfunction has to be understood
in a very broad sense (e.g., inclusive of human capital), so that the
estimates obtained conform to the expected share of such broadly
defined capital in output. M-R-W, however, suggested explicit
inclusion of human capital as another input of the production
function and hence as a variable in the regression equation. They
showed that this augmentation by human capital leads not only to
a better fit of the model, but also to more realistic estimates of a.
M-R-Wconsideredout of steady state behavior as well. They noted
that the countries may not be in their steady states (or the
departuresfrom steady states may not be randomacross countries)
and hence proceededto see how far the (augmented) Solow model
proves successful in describingthe transitional dynamics.
In both of these exercises, M-R-Wrelied on a crucial assumption. Apart from the saving and population growth variables,
equation (3) contains the term [ln A(O) + gt)]. Since the exogenous
rate of technological progress, g, is thought to be the same for all
countries and for a cross-sectionregression t is just a fixed number,
gt in the equation is just a constant. However,this cannot be said of
A(O). M-R-W rightly noted, "the A(O) term reflects not just
technology but resource endowments, climate, institutions, and so
on; it may therefore differacross countries" [p. 6]. They, therefore,
postulated that
lnA(O) = a +
E,
where a is a constant and e is the country-specificshift or shock

term. Substituting this into the equation above and subsuming gt

1134
QUARTERLY JOURNAL OF ECONOMICS
into the constant term a, they derived the specification:

(4)
My
In \L=a
aa
+ 1_
In (s
1-(
In (n
+g
+ 8)
+ e.
At this stage, however, M-R-W made the assumption that E is

independent of the explanatory variables, s and n. This was their
identifying assumption, and this allowed them to proceed with the
Ordinary Least Squares (OLS) estimation of the equation. M-R-W
provided several arguments for this assumption. The first is that
this assumption is common and is made not only in the Solow
model, but also in other growth models. Also, they noted that in
models where saving and population growth are endogenous but
preferences are isoelastic, s and n are independent of E. Second, this
identifying assumption renders it possible to test various informal
hypotheses that have been made (proceeding from different growth
theories) regarding the relationship between income, saving, and
population growth. Third, since the specification above postulates
not only the signs of the coefficients but also their proximate
magnitudes, the regression results will allow testing of the joint
hypothesis of validity of the Solow model and the above-mentioned
identifying assumption.
Of these arguments the most important one is the first.
However, the assumption of isoelastic preference represents an
additional restriction. In general, the country-specific technology
shift term E is likely to be correlated with the saving and population
growth rates experienced by that country. At a heuristic level, since
A(O) is defined not only in the narrow sense of production
technology, but also to include resource endowments, institutions,
etc., it is not entirely convincing to argue that saving and fertility
behavior will not be affected by all that is included in A(O).
What is important to note here is that in the framework of a
single cross-section regression, this assumption of independence
becomes an econometric necessity. OLS estimation is valid only
under this assumption. The other possibility in this regard is to
recognize the correlation and then opt for instrumental variable
(IV) estimation. However, given the nature and scope of the A(O)
term, it is difficult to come up with instruments that will be
correlated with the included explanatory variables of the model and
yet uncorrelated with A(O). This makes the option of instrumental
variable estimation not quite feasible.
Our basic conjecture is that a panel data framework provides a
better and more natural setting to control for this technology shift

1135
term E.This is better revealed by considering the equation describing out of steady state behavior. The way this equation is derived is
as follows. Let 9* be the steady state level of income per effective
worker, and let 9(t) be its actual value at any time t. Approximating
around the steady state, the pace of convergence is given by
dIn9(t)
(5)
9(t)],
X[ln (9*)-In
dt-
dt
where A = (n
+ g + 8)(1 - oa).This equation implies that
(6)
In9(t2)
In9* + e-AT In9(t1),

where 9(t1) is income per effective worker at some initial point of
time and T = (t2 - t1). Subtracting In 9(t1) from both sides yields
(7)
(1
e-AT)
1n9(t2) - In9(t1) = (1 - e-At) In 9* - (1 -
e-AT)
In9(t1).
This equation represents a partial adjustment process that becomes more apparent from the following rearrangement:
In9(t2) - In 9(t1) = (1 - e- XT)(In9* - In9(t1)).
(8)
In the standard partial adjustment model, the "optimal" or

"target" value of the dependent variable is determined by the
explanatory variables of the current period. In the present case, 9*
is determined by s and n, which are assumed to be constant for the
entire intervening time period between t1 and t2 and hence
represent the values for the current year as well. Substituting for
9* gives
(9)
1n9(t2) - In 9(t1) = (1 - e -AT)
-(1
e XT)
ln
1-(x
In (s)
(n +g+5)
- (1 -e-XT)ln9(tl).
M-R-W used this equation to study the process of convergence

across different samples of countries. In their treatment t1 was
1960, and t2 was 1985. They assumed (g + 8) to be the same for all
countries and equal to 0.05. The saving and population growth
rates, s and n, were taken to be equal to the respective averages
over 1960-1985.
The issue of correlation between the unobservable A(0) and
the observed included variables is not apparent in equation (9)
because it has been formulated in terms of income per effective
worker. In actual implementation, however, M-R-W worked with

1136
income per capita. We may, therefore, reformulate the equation in

terms of income per capita. Note that income per effective labor is
Y(t)
L(t)A(t)egt
Y(t)
9(t)
A(t)L(t)
so that
IY(t)\
I - In A(O) - gt
In9(t) = In I\
L(t)/
= In y(t) - In A(O) - gt,
where y(t) is the per capita income, [Y(t)/L(t)]. Substituting for 9(t)
into equation (9), we get the usual "growth-initial level" equation:
(10)
Iny(t2) - Iny(t1) = (1 - e-T)

-(1
n (s)
I_
l (n + g + ) -(1
In
- e-AT)
+ (1 - e
- e-AT)lny(t1)
AT) In A(O) +
g(t2 -e
-Tt)
However, if we collect terms with In y(t1) on the right-hand side, we

get the equation in the following alternative form:
(11)
lny(t2) = (1 - e-AT) 1
- (1
e-AT)
ln (s)
+g + 8) + e -Tlny(t1)
aIn(n
+ (1 - e
AT)
InA(O) + g(t2 -e
-Ttl
It can now be seen that the above represents a dynamic panel

data model with (1 - e -T) In A(O) as the time-invariant individual
country-effect term. We may use the following conventional notation of the panel data literature:
2
(12)
Yit = YYit-i + E jijxI't+ nt + pi + vit,

j=1
where
yit=
yist-l =
y=
lny(t2)
lny(t1)
e-XT

e~
-(
1137
-XeT)
I2 = -(1 - e-AT)
1-
il = In (s)
x2= In (n + g +)
ki = (1 It
= g(t2
e-AT)
-
In A(0)
XTt )I
and vit is the transitory error term that varies across countries and
time periods and has mean equal to zero. Panel data estimation of
this equation now provides the kind of environment necessary to
control for the individual country effect.
It is clear that this panel data formulation is obtained by
moving from a single cross-section spanning the entire period
(1960-1985) to cross sections for the several shorter periods that
constitute it. We may note that equation (11) was based on
approximation around the steady state and was supposed to
capture the dynamics toward the steady state. It is, therefore, valid
for shorter periods as well. Also, it may be noted that in the single
cross-section regression, s and n are assumed to be constant for the
entire period. Such an approximation is more realistic over shorter
periods of time. The panel data setup allows us, after controlling
for the individual country effects, to integrate this process of
convergence occurring over several consecutive time intervals. If
we think that the character of the process of getting near to the
steady state remains essentially unchanged over the period as a
whole, then considering that process in consecutive shorter time
spans should reflect the same dynamics. However, controlling for
the unobservable individual country effects will create a cleaner
canvas for the relationship among the measurable and included
economic variables to emerge.
IV. ESTIMATION
ISSUESANDDATA
A. Relevant Issues of Panel Estimation
A host of methods is available for the estimation of panel data
models with individual effects.1 One common issue that arises in
1. For a recent review see Islam [1991].

1138
such estimation is whether the individualeffects are to be thought

of as "fixed" or "random." In the latter case the effects are
assumed to be uncorrelatedwith the exogenous variables included
in the model. In our case it is clear that estimators relying on such
assumptions (for example, GLS in Maddala[1971]) are not suitable
because it is preciselythe fact of correlationthat forms the basis of
our argumentation for the panel approach.
The Least Squares with Dummy Variables (LSDV) estimator
which is based on the fixed-effectsassumption is still permissible,
although that assumption may seem too strong. One problemwith
LSDV for our model, however, arises from its dynamic character.
The presence of a lagged dependentvariableon the right-hand side
of equation (12) makes LSDV an inconsistent estimator, when
asymptotics are considered in the direction of N
->
oo. However, the
asymptoticpropertiesof panel data estimators can be consideredin

the direction of T, and Amemiya [1967] has shown that when
considered in that direction, LSDV proves to be consistent and
asymptotically equivalent to the Maximum Likelihood Estimator
(MLE). Yet many other estimators start by eliminating the individual-effectterm through first differencing.For these estimators,
therefore, it does not matter whether the effect is fixed or random,
and in the latter case, whether correlatedor not.
The theoretical properties of most of these estimators are
asymptotic,and in terms of these propertiesthey are equivalent. In
order to decide which of these to use, we conducted a Monte Carlo
study based on the data set used in this paperand closely calibrated
to the parameter values obtained from preliminary estimation.
These results show that the LSDV estimator, although consistent
in the direction of T only, actually performs very well. The other
estimator that showed better performance is the Minimum Distance (MD)estimator proposedby Chamberlain[1982, 1983].2 This
estimator is specially designed for models where the individual
effects are correlated with the included exogenous variables. The
MD estimator has the addedattractive propertythat it is robust to
any presence of serial correlation in the vitterm. In the following
we present the results from both LSDV and MD estimation. It is
reassuringthat the results are very similar to each other.
B. Data and Samples
One of the reasons for the recent surge in work on growth

empirics has been the availability of the Summers-Heston [1988]
2. Details of these results can be seen in Islam [1992].

1139
data set. In fact, the emergence of new (endogenous) growth

theories can be traced to some of the empirical regularities of
cross-countrygrowth which this data set helped to bring to the fore
and which at first sight seemed to contradictthe implicationsof the
existing neoclassical growth models. It is interesting to note that
the Summers-Heston data set also makes a panel approach to
growth empirics possible because it includes various measures of
the GDP and its components for different countries over several
decades. Both Barro [1989] and M-R-Wused the Summers-Heston
data set to construct the variables. Our present exercise is also
based on this data set.
Since we want to compareour panel results with, in particular,
those obtained by M-R-Wfrom their single cross-section regressions, we keep the country samples similar to those used by them.
The three samples that M-R-Wconsidered were (i) NONOIL (98
countries), (ii) INTER (75 countries), and (iii) OECD (22 countries). However, data for some of the initial years on Indonesia and
Burkina Faso are not available in the Summers-Heston data set.
We therefore excludedthese two countries from our samples. This
led to our NONOIL sample having 96 countries, while the INTER
sample has 74 countries. The size of the OECDsample remains the
same.
The other respect in which our variable construction differs
from that of M-R-Wis in the treatment of the population growth
variable n. M-R-Wtook n as the rate of growth of the working age
population. In view of the difficulty of getting panel data on
working age population,we dependon the populationgrowth rates
computed from the total population figures available in the Summers-Heston data set. However, following M-R-W,we take (g + 6)
to be equal to 0.05 and assume this value to be the same for all
countries and all years.
The switch from a single cross section to a panel frameworkis
made possible by dividingthe total periodinto several shorter time
spans. The question that arises is what is the appropriatelength of
such time spans. The furthest that one can go in this regard is to
consider a time span of just one year (which is technically feasible
given that the underlying data set provides annual data). For
several reasons, however, it seems that yearly time spans are too
short to be appropriatefor studying growth convergence. Shortterm disturbances may loom large in such brief time spans.
Instead, we opt for five-yeartime intervals. Thus, consideringthe
period 1960-1985, we have five data (time) points for each country:
1985, 1980, 1975, 1970, and 1965. When t = 1965, for example, t -

1140
1 is 1960, and saving and population growth variables are averages

over 1960-1965. With this setup, the vi,'s are now five calendar
years apart (alternatively, pertain to five-year spans) and hence
may be thought to be less influenced by business cycle fluctuations
and less likely to be serially correlated than they would be in a
yearly data setup.
V. ESTIMATION
RESULTS
A. Single Cross-Section Results
In order to see how much our results differ from those of
M-R-W because of differences in samples and construction of
variables, we first run single cross-section regressions analogous to
those conducted by M-R-W. For these regressions Yit is the log of
per capita GDP for 1985 and yit-1, the same for 1960. s and n are
averages of saving and population growth rates for the period
1960-1985. The results can be seen in Table I. The first panel of
the table gives results of estimation in unrestricted form, while the
second panel contains results from estimation of the equation after
imposing the restriction that the coefficients of the investment and
population growth variables are equal in magnitude but opposite in
sign. M-R-W's results (their Table IV) on this regression are
available only in the unrestricted form. We can therefore compare
the unrestricted results only.
Such a comparison shows that the results are very similar. The
coefficients of the initial GDP and saving variable are very close to
each other in the two tables. Modified for the difference in the way
the equation is specified, our estimates of the initial GDP variable
for NONOIL, INTER, and OECD samples will be -0.127, -0.218
and -0.328, respectively. Corresponding estimates of M-R-W are
-0.141, -0.228, and 0.351, respectively. This is also reflected in
the respective implied values of the rate of convergence parameter
X. Our values of 0.00542, 0.0098, and 0.0 159 for NONOIL, INTER,
and OECD are very close to the corresponding M-R-W estimates,
namely, 0.00606, 0.0104, and 0.0173, respectively.
The results from restricted estimation allow us to get unique
estimates of not only X, but also the output elasticity parameter, a.
The estimates of X obtained from restricted estimation are almost
the same as those from unrestricted estimation. In general, they
confirm the finding of a very slow rate of convergence. On the other
hand, the estimate of a is found to be 0.83 for the NONOIL sample,
0.76 for INTER, and 0.60 for OECD. These are, indeed, unusually

1141

TABLE I
SINGLE CROSS-SECTIONRESULTS, 1960-1985:
DEPENDENT VARIABLEIs
NONOIL(96)
Sample:
INTER(74)
In (y85)
OECD(22)
Unrestricted
Constant
In (Y60)
In (s)
In (n + g + 5)
0.9448
1.1075
1.7433
(0.8724)
0.8733
(0.0611)
0.6585
(0.0926)
(0.8975)
0.7822
(0.0667)
0.6431
(0.1121)
(1.2655)
0.6722
(0.0694)
0.4114
(0.1845)
-0.6122
(0.3667)
0.9006
0.00542
(0.00037)
R
ImpliedX
-0.8144
(0.3717)
0.8915
0.009827
(0.00083)
-0.8021
(0.4187)
0.8499
0.015887
(0.00164)
Restricted
Constant
In (Y60)
In (s) - In (n + g + 5)
R
Implied X
Implied a
Wald test of restriction:

p-value
0.8475
1.4565
2.6689
(0.3429)
0.8701
(0.0547)
(0.3798)
0.7945
(0.0599)
(0.5715)
0.6817
(0.0678)
0.6554
0.6610
0.4847
(0.0884)
0.9037
0.005565
(0.00035)
0.8346
(0.1126)
(0.1034)
0.8927
0.009204
(0.00069)
0.7628
(0.1193)
(0.1602)
0.8524
0.015327
(0.00152)
0.6036
(0.1995)
0.95
0.90
0.70
Figures in parentheses are standard errors.
high values for an elasticity of output with respect to capital.

M-R-W's corresponding estimates are not available. However, we
know from their discussion that it was these high estimates of a
that led them to suggest inclusion of human capital as another
factor of the production function.
B. Pooled Estimation
We next see whether dividing the growth period into five-year
spans has any significant effect. For this we implement a pooled
regression (OLS) on the basis of our five-year span data. The
results from such estimation can be seen in Table II.
It is striking to note how similar these results are to those

1142
TABLE II
POOLED REGRESSIONFROMA PANEL OF FIvE-YEAR SPAN DATA:
DEPENDENT VARIABLEIs Yit
NONOIL
480
Sample:
No. of obs.
INTER
370
OECD
110
0.9636
(0.0107)
0.1396
(0.0172)
-0.1300
(0.0566)
0.9861
0.0074
(0.0001)
0.9228
(0.0147)
0.1047
(0.0313)
-0.1799
(0.0653)
0.9807
0.0161
(0.0003)
0.9628
(0.0098)
0.1388
(0.0165)
0.9861
0.0095
(0.0002)
0.7736
(0.0924)
0.9248
(0.0147)
0.1184
(0.0286)
0.9901
0.0146
(0.0002)
0.6150
(0.1486)
Unrestricted
0.9764
(0.0101)
0.1386
(0.0153)
-0.1291
(0.0584)
0.9848
0.0048
(0.0001)
in (yi,t-i)
in (s)
in (n + g +5)
R2
Implied X
Restricted
in (yi,t-1)
in (s)-in
(n + g +5)
R2
Implied X
Implied a
Wald test for restriction:

p-value
0.9758
(0.0012)
0.1381
(0.0151)
0.9848
0.0059
(0.0001)
0.8338
(0.0912)
0.90
0.90
0.90
obtainedfrom the single cross section. Because of differencesin the

value of T, the reduced-formcoefficients are not directly comparable. We, therefore, look at the implied values of the structural
parameters. The values of X from the unrestricted pooled estimation are 0.0048, 0.0074, and 0.0161 for NONOIL, INTER, and
OECD. These are quite close to the corresponding estimates in
Table I. It is only with the INTER sample that we notice some
sizable difference. However, even this difference disappearswhen
we comparethe estimates from restricted regression. The implied
values of . obtainedfrom pooledregression are 0.0059, 0.0095, and
0.0146 for the NONOIL,INTER, and OECDsamples, respectively.
These are almost the same as those reportedin the second panel of

GROWTH EMPIRICS: A PANEL DATA APPROACH
1143
Table I. The implied values of a obtained from these two regressions are also found to be strikingly similar.
These results, therefore, show that dividing the period into
shorter spans and considering the growth process over shorter
consecutive intervals does not affect the results. Both the single
cross section and the pooled regression produce very similar
results. We find very low estimates of the rate of convergence
(particularlyfor the NONOIL and INTER samples) and very high
estimates of the elasticity parameter a. We next see how panel
estimation changes these results.
C. Minimum Distance: Estimation with "Correlated Effects"
The MD estimator emphasizes the correlated nature of the

individual-effectterm and does not eliminate it by differencingthe
equation. Instead, it attempts to incorporatethe correlationin the
estimation process by explicitly specifying kLias a function of the
variableswith which it is thought to be correlated.
One simple specificationof kLisuggested by Mundlak [1978] is
to take it as a function of the mean of the exogenous variable
pertaining to the individual, xi (assuming that we have only one
exogenous variable, xit, in the model and 3 is its coefficient).
Mundlak's purpose in using such a simple specification was,
however, to show that if kLiis a linear function of x-i,then the GLS
estimation under the random effects assumption reduces to the
LSDVestimation under the fixed effects assumption.
Chamberlain noted that the specification of ,i suggested by
Mundlak was overly restrictive. He, therefore, proposed a more
general specification whereby kLidepends linearly on the xi for all
time periods. Thus, we would have
(13)
,ui = Ko + K1Xij
with E [i lxi,, . .
., XiT]
+ KTXiT + PJi
+ K2Xi2 +
0. Note that, regarded as a linear
predictor, the above specification does not entail any restriction.

This then allows substitution for kLi in equation (12) by its
specification in terms of xit's as given by equation (13). Also, by
repeated substitution we can replacethe lagged dependentvariable
on the right-handside by expressions involvingyi0,the initial value.
Instead of assuming the yiO's as given and fixed, Chamberlain
suggested a similar general specificationforyi in terms of xit's:
(14)
Yio = No + (lXil
+ (2Xj2 +
...
+ (TXiT + Pi,

1144
where, again, E[Rjlxij.... XiT1] - 0. Note that, interpreted as

linear predictor,the above specificationis perfectlygeneral. We can
then substitute equation (14) for Yioin equation (12). Through
these substitutions we arrive at a system of reduced-formequations, all of which have only the xi,'s as the right-hand side
variables. Estimation of these reduced-formequations gives us the
HI-matrixof the reduced-formcoefficients. Structural parameters
are then obtained by imposing the nonlinear constraints on the
IH-matrixthrough the MD procedure. A brief discussion of this
estimation procedurehas been providedin Appendix4.
We implement the MD estimation procedurefor the equation
in the restricted form. Hence the equation is
,
(15)
YU =
Yist-l +
exit
mt+ ki
+ Vit,
where
ax
=
(1
e-XT)
xit = In (s) - In (n + g + a).
The rest of the notation is as in equation (12). Since the MD results

showed some sensitivity to the starting values of the iteration
process, we conducted a sensitivity analysis based on the MD
statistic and obtained results that yielded the minimum value of
this statistic for different alternative combinations of the starting
values for -yand P. In other words, we tried to achieve the global
minimum instead of a local one. However,it needs to be mentioned
that due to small N, the MD estimates for the OECDsample may be
less accuratethan for the other two samples.3
The results from MD estimation can be seen in Table III. For
reasons stated earlier, we again focus on the estimated values of
the structural parametersonly. It is clear from Table III that panel
estimation allowing for correlated individual country effects leads
to considerablechange in the results. The implied value of Xfrom
the MD estimation for the NONOILsample is 0.0434, for INTER it
is 0.0456, and for OECD, 0.0670. These are indeed much higher
values than the correspondingsingle cross-sectionvalues. The rate
of convergenceobtained for the OECD sample still remains much
higher than that found for either the NONOILor INTER samples,
but even for these latter samples it is now appreciablyhigher. In
3. The weighting matrix becomes nearly singular, and sensitivity analysis with
respect to the initial values for the parameter ,3 remains less than conclusive.

1145
TABLE III
MINIMUM DISTANCE ESTIMATIONWITH CORRELATEDEFFECTS:
DEPENDENT VARIABLEIS yit
Sample
y
0
Implied X
Implied at
NONOIL
INTER
OECD
0.8050
(0.0306)
0.1530
(0.0274)
0.0434
(0.0076)
0.4397
(0.0614)
0.8117
(0.0284)
0.1389
(0.0243)
0.0417
(0.0070)
0.4245
(0.0524)
0.7155
(0.0098)
0.1203
(0.0264)
0.0670
(0.0026)
0.2972
(0.0433)
Figures in parentheses are asymptotic standard errors.
fact, the relative increase in the impliedvalue of Ais the highest for
the NONOIL sample (by a factor of about 8). Another change that
results from panel estimation is that the impliedvalues of Xfor the
NONOIL and INTER samples are closer to each other. While on
the basis of the single cross-section results, the implied value of A
for INTER was almost twice that for NONOIL, corresponding
values from MD estimation do not differby that much.
Significant change is also observedin the implied values of the
elasticity parameter, at.The estimated value of a for the NONOIL
sample is 0.4397, for INTER, 0.4245, and for OECD,0.2972. These
may be contrasted with the corresponding single cross-section
estimates: 0.8346, 0.7628, 0.6036, respectively. The across-sample
pattern of variation of the parameter estimates remains the same
as obtainedfrom single cross-section regression. The impliedvalue
of the elasticity parameter is found to be the highest for the
NONOIL sample and lowest for the OECD sample. But the panel
estimates for all the samples are much lower and much closer to
their generally accepted values. It is also interesting to note that
the estimated values of a that we obtain from panel estimation are
very close to the estimates that M-R-Wobtainedafter includingthe
human capital variable.
D. LSDV:Estimation with "FixedEffects"
Very similar results are obtained even if one assumes that the
individualcountry effects are fixed in nature. This can be seen from
the results of the LSDV estimation presented in Table IV. Again,
focusing on the estimates of the structural parameters,we see that

1146
TABLE IV
LSDV ESTIMATIONWITH FIXED EFFECTS: DEPENDENT VARIABLEIs Yit
Sample:
No. of obs.
Unrestricted
Yit-1
In (s)
In (n + g + 8)
R2
ImpliedX
Restricted
Yit-1
In (s) - In (n + g + 8)
R2
ImpliedX
Implied
Wald test for restriction:

p-value
NONOIL
480
INTER
370
OECD
110
0.7762
(0.0353)
0.1595
(0.0237)
-0.4092
(0.1024)
0.7404
0.0507
(0.0091)
0.7935
(0.0388)
0.1709
(0.0256)
-0.2466
(0.1007)
0.8254
0.0462
(0.0098)
0.5864
(0.0532)
0.1215
(0.0586)
-0.0698
(0.1007)
0.9659
0.1067
(0.0181)
0.7919
(0.0349)
0.1634
(0.0238)
0.7368
0.0467
(0.0088)
0.4398
(0.0545)
0.7954
(0.0387)
0.1726
(0.0254)
0.8251
0.0458
(0.0097)
0.4575
(0.0575)
0.6294
(0.0495)
0.0954
(0.0581)
0.9642
0.0926
(0.0157)
0.2047
(0.1042)
0.70
0.90
0.90
Figures in the parentheses are standard errors, and the goodness of fit measures are with respect to the
within regression.
the implied rates of convergence for the NONOIL, INTER, and

OECD samples are 0.0467, 0.0458, and 0.0926, respectively. The
corresponding estimates of the output elasticity with respect to
capital are 0.4398, 0.4575, and 0.2047. These are remarkably
similar to the correspondingestimates obtained from MD estimation. The similarity is particularly striking for the NONOIL and
INTER samples. The LSDV estimates for the OECD sample differ
to some extent from the correspondingMD estimates. However,
the direction in which they differ accentuates the qualitative
properties of panel estimation results. The implied rate of convergence for the OECD sample obtained from LSDV estimation is
even higher, and the output elasticity lower than those obtained
from the MD estimation.
In sum, therefore, adoption of the panel approach leads to a

1147
twofold change in the results. First, we obtain much higher rates of

convergence, and second, we obtain more empirically plausible
estimates of the elasticity of output with respect to capital. In the
section below we try to considerthe source and implicationsof each
of these results.
VI. INTERPRETATION OF THE RESULTS
A. Statistical Interpretation
The statistical source of the change in the parameterestimates

is not very difficult to comprehend. Both of the above-noted
changes can, to a great extent, be attributed to correction for
omitted variablebias that the panel approachmakes possible.
We have seen in Section II that, in the framework of single
cross-section regression, the A(O) term, being unobservable or
unmeasurable,is left out of the equation (or, subsumed in the error
term). This actually creates an omitted variableproblem.Since this
omitted variable is correlated with the included explanatory variables, it causes the estimates of the coefficientsof these variablesto
be biased. The direction of bias can be assessed from the standard
formula for omitted variable bias. The partial correlationbetween
A(O) and the initial value of y is likely to be positive, and the
expected sign of the A(O)term in the full regression, as can be seen
in equation (11), is also positive. Thus, 'y,the estimated coefficient
ofyi t_1,is biased upward.The relationshipbetween Aand y is given
by
(16)
A = (1/X) In (,y).
leads to a lower value

This equation shows that a higher value of My
of X.This explains why we get lower convergencerates from single
cross-section regressions and pooled regressions that ignore correlated individualcountry effects.
Similarly, the relationship between axand the reduced-form
coefficients y and 3is
(17)
a = /(1-+
13).
This formula for a clearly shows that overestimation of y also leads

to a higher implied value of at. Since 3 figures in both the
numerator and the denominator, the net effect of any bias in the
estimate of Pon the value of a is not immediately clear. It depends
on the range of values that this estimated coefficient takes. We

1148
have seen above what the likely effect on the estimate of y can be
when the correlatedindividualcountry effect is ignored.
We may apply similar reasoning to assess the likely effect of
ignoring the individual effect on the estimate of P. Recall that in
the restricted form, P3is the coefficient of [In (s) - In (n + g + 6)].
The theoretical sign of the coefficient is positive. However, it is
difficult to be sure about the partial correlation of A(O) with
[In (s) - In (n + g + 6)]. Whatever may be the direction of bias in
the estimate of P3,because of its presence in both the numerator
and the denominator,the effect of this bias may, to a large extent,
cancel out, and bias (and its correction)in the estimate of y may be
a determining factor. As equation (18) shows, a lower value of 'y
leads to a lower value of a.
This is, obviously, the narrow statistical interpretation of our
results. We next turn to the interpretation and implications of the
results from growth theory's point of view.
B. Estimation Results and Growth Theory
In the growth literature of recent years, three empirical

results have surfaced. These are (i) absence of absolute convergence among the countries in the larger sample, (ii) slow conditional convergenceamong countries in the larger sample, and (iii)
absolute or faster conditional convergence among "similar" subgroups of countries of the larger sample. As we have previously
noted, it was the first finding (with or without the third) that
triggeredthe developmentof new (endogenous)theories of growth.
The concept of conditional convergenceto some extent reinstated
the "old" Solow-Cass-Koopmanstheory of growth, although, so
far, that needed incorporation of human capital into the model.
However, even after accounting for human capital, growth empirics based on a single cross-section regression yielded a rather slow
rate of conditional convergence, which is problematic,because an
open economyversion of the Solow-Cass-Koopmansmodel predicts
instantaneous convergence.
It is in this context that we need to evaluate the panel
estimation results that we have obtained. First of all, the finding of
a faster rate of conditional convergence (even without taking
account of human capital) is obviously good news for the SolowCass-Koopmansmodel. The rates of conditional convergence obtained from panel estimation are nowhere near infinity, but they
are much higher than the correspondingrates obtainedfrom single
cross-section regressions and hence lend somewhat more validity

1149
to the cross-country implications of the Solow-Cass-Koopmans

model.
Equation (3) shows that the steady state income levels differ
across countries not only because of differences in s and n (and
possibly because of differences in 8, and, in particular, g), but also
because of differences in A(O). The question is whether differences
in A(O) are important enough to have significant impact on
cross-country growth regularities. If they were not important,
holding these constant through the panel framework, would not
make much difference in the results. The fact that it does shows
that these differences indeed play an important role in understanding the international growth experience. In other words, the A(O)
term is an important source of parametric difference in the
aggregate production function across countries. The process of
convergence is thwarted to a great extent by persistent differences
in technology level and institutions.
This finding is consistent with the generic finding of faster
convergence among groups of similar countries that have been
reported earlier by researchers. Instead of adopting the panel data
approach, the other way to control for differences in technology and
institutions is to classify the countries into similar groups. Baumol
[1986] coined the term "convergence club" to express this phenomenon. The classification itself can, however, be problematic. The way
Baumol did it suffered from self-selection bias, as was demonstrated by De Long [1988]. Recently, Chua [1992] did an analysis
where similar could be interpreted as geographic contiguity. Durlauf
and Johnson [1991] attempted to endogenize the classification
using the "regression tree" method. In all these cases, convergence
was found to be much stronger within the groups and weak
between them. Durlauf and Johnson were quite emphatic about
the cause of this result. They concluded that the aggregate
production function differed across different locally convergent
groups and hence suggested that, ". . . the Solow growth model
should be supplemented with a theory of aggregate production
function differences in order to fully explain international growth
patterns" [p. 1]. What we have done in this paper is, by adoption of
a panel data approach, to allow for differences in the aggregate
production function not only across groups of countries (however
defined), but across individual countries. As a result, we obtain
higher rates of convergence over the samples as a whole.
Having seen the impact of inclusion of individual country
effects on growth regression results, we now turn to the question of

1150
what happens when human capital is brought into the panel

frameworkof analysis.
VII. INCLUSIONOF HUMAN CAPITAL
Measures of human capital have always been a weak spot in

growth empirics. In fact, M-R-Wprovide a very good discussion of
the problemsand issues involved in this regard.We cannot use the
identical SCHOOL variable in our analysis because analogous
panel data on this variable are difficultto come by. Moreover,since
M-R-Ws'work, Barroand Lee [1993] have made importantprogress
in putting together a human capitaldata set for a wide cross section
of countries. Based on census data and myriads of other information they have constructed a human capital variable, named
HUMAN, which gives the average schooling years in the total
population over age 25. While the SCHOOLvariable is based on
secondaryschooling information only, HUMANincludes schooling
at all levels, primary, secondary, and higher, complete and incomplete. Second, HUMAN gives a direct measure of the stock of
human capital, and hence makes it possible to estimate the
equation in which human capital appearsas a stock. The restricted
form of this equation is
(18) lny(t2) = (1 + (1
e-AT)
eXT)
ax
[In (s) - In (n + g + 8)]
In (h*) + e -T In y(tl)
+ (1 - e
AT)
In A(0) + g(t2 -
eXTti)
where h* is the steady state level of human capital, and (pis the
exponent of the human capital variable in the augmented production function of M-R-W.
Inclusion of the human capital variable in our analysis,
however, requires some modification of the sample sizes. This is
because the panel data on HUMANin Barroand Lee [1993] are not
availablefor all the countries of our original samples. Accordingly,
the NONOIL sample now reduces to 79 countries,4and INTER to
67.5 Size of the OECDsample, expectedly,remains unchanged.
4. The countries dropped from the original NONOIL sample are Angola,
Benin, Burundi, Cameroon, Central African Republic, Chad, Congo, Egypt, Ethiopia, Ivory Coast, Madagascar, Mali, Mauritania, Morocco, Nigeria, Rwanda, and
Somalia.
5. The countries dropped from the original INTER sample are Cameroon,
Ethiopia, Ivory Coast, Madagascar, Mali, Morocco, and Nigeria.

1151
TABLE V
ESTIMATIONWITH HUMAN CAPITAL
Variable
Single
cross section
Pooled
regression
Panel
estimation
0.0093
(0.0146)
0.0069
(0.0025)
0.8013
(0.0534)
0.0544
(0.1020)
-0.0712
(0.0323)
0.0375
(0.0093)
0.5224
(0.0642)
-0.1990
(0.1097)
-0.0014
(0.0209)
0.0079
(0.0028)
0.7854
(0.0587)
-0.0077
(0.1288)
-0.0027
(0.0471)
0.0444
(0.0102)
0.4947
(0.0599)
-0.0069
(0.1261)
0.0034
(0.0268)
0.0162
(0.0055)
0.6016
(0.1015)
0.0174
(0.1797)
-0.0208
(0.0449)
0.0913
(0.0160)
0.2074
(0.1055)
-0.0450
(0.1457)
NONOIL
ln(h)
Implied X
Implied at
Implied up
0.1823
(0.0895)
0.0111
(0.0038)
0.6862
(0.0694)
0.2356
(0.1013)
INTER
ln(h)
Implied X
Implied a
Implied up
0.1101
(0.1305)
0.0118
(0.0045)
0.6906
(0.0799)
0.1335
(0.1428)
OECD
ln(h)
Implied X
Implied ax
Implied up
0.0864
(0.1551)
0.0187
(0.0077)
0.5416
(0.1426)
0.1062
(0.2027)
Before moving on to panel estimation, it is instructive to note

the results obtained from inclusion of the human capital variable in
the single cross-section and pooled regression frameworks.6 These
have been presented along with results from panel estimation in
Table V. Looking at the single cross-section results, we observe the
following. First, in general, the outcome is similar in spirit to that
found by M-R-W and other researchers. Inclusion of the human
6. We also check for the effect of the modification of the samples on estimation
without human capital. These results are not presented here. In general, the sample
modifications do not significantly change the results.

1152
capital variable in the single cross-section regression framework

does lead to higher rates of convergence and lower values of at.In
our case, however, the changes are of a lesser order.7 Second,
judging by the size of the standard errors, the human capital
variable does not prove to be significant for all the different
samples. Even in M-R-Wthe SCHOOLvariabledid not prove to be
significant for the OECDsample. In our case the same is found for
the INTER sample as well. Third, while for the NONOIL sample
the value of (p(0.2356), the implied exponent for the human capital
variable, was found to be very similar to that found by M-R-W,for
the INTER and OECD samples the estimates are substantially
lower: 0.1335 and 0.1062, respectively, compared with 0.23 of
M-R-Win both cases.
Turning to pooled estimation, we find that the results change
in a very different direction. First of all, we note that the human
capital variable now loses statistical significance for the NONOIL
sample as well. Second, for the INTER sample it even assumes the
wrong sign. Third, with respect to the impact on the impliedvalues
of X,the estimates now decrease to 0.0069, 0.0079, and 0.0162 for
the NONOIL, INTER, and OECD samples, respectively. On the
other hand, estimated values of axincrease to 0.8013, 0.7854, and
0.6016, respectively. These estimates are very similar to those
obtained from single cross-section regression without the human
capital variable.What this implies is that incorporationof the time
dimension of the human capital variable into the analysis annihilates the effect that the cross-sectionalvariation in human capital
had on the regression results. It is against this backgroundthat we
now consider the results from panel estimation. These can be seen
in the last column of Table V.
The figures in that column show that introduction of individual country effects reproduces the earlier result even in the
presence of time series of human capital. The implied values of X
are now found to be 0.0375, 0.0444, and 0.0913 for the NONOIL,
INTER, and OECD samples, respectively. The implied values of ax
are 0.5224, 0.4947, and 0.2074 for these samples, respectively.
Note that these results are broadlysimilar to the panel results that
7. There may be several reasons for this. First, the samples are not the same.
Second, the human capital variable used is not the same. Third, the form in which
the human capital variable has been entered into the equation is different. Equation
(19) requires a steady state level of human capital. In actual implementation we
have used the HUMAN of the end points of time for the respective time spans. This
is different from using either the initial measure of human capital (as in Barro
[1989]) or the average for the period (as in M-R-W [1992]).

1153
we obtained earlier without including human capital. This is also

not surprising in view of the fact that the human capital variable
does not proveto be significantfor two of the three samples. Also of
note is the fact that in all three samples, the coefficient on the
human capital variable now appears (in the restricted version of
the model) with the wrong sign. In the case of the INTER sample,
the magnitude of the coefficient and the implied value of .p are
negligible. To a great extent, the same is true for OECDsample. It
is only for the NONOILsample that we find the negative coefficient
on the human capital to be sizable and marginallysignificant.
However, such "anomalous" results regarding the role of
human capital in the growth process are not new. Whenever
researchershave attempted to incorporatethe temporal dimension
of human capital variables into growth regressions, outcomes of
either statistical insignificanceor negative sign have surfaced.8So
far, there have been two kinds of responses to these types of
results. One is to point out the discrepancybetween the theoretical
variableH in the productionfunction and the actual variable used
in regressions. The enrollment rates were always very partial
measures of the rate of investment in human capital and, more
importantly, did not account for differences in the quality of
schooling. Although, measured by such rates, many (particularly
the less developed)countries appear to have made much progress,
the true levels of human capital (and hence the output levels) in
these countries have actually not increased by that much. Statistically this results in a negative temporal relationship between the
human capital variable used and economic growth within countries. The results of the pooled regression already show that this
negative temporal relationship is strong enough to outweigh the
positive cross-sectional relationship. Since panel estimators rely
more on "within" variation, we find that in panel results the
negative temporal relationship surfaces more forcefully. From this
point of view, what our results show is that even Barro and Lee's
HUMAN variable, though much more comprehensivein its scope,
is nonetheless not free from the above issue of discrepancy.Also, in
light of the above, it is no wonder that it is in the NONOIL sample
that the negative effect of the HUMANvariableis found to be more
pronounced.
The second response is to think of richer specification of the
production function with respect to human capital. In a sense,
8. For one such example see De Gregorio [1991].

1154
Romer [1989b] can be thought of as a pioneer in this regard. The

recent work by Benhabib and Spiegel [1994] is another significant
step in this direction. Their estimation of a growth equation in the
first differencedform (which may be regardedas panel estimation
with two periods only), results in insignificant or negative coefficients on the human capitalvariablefor all differentsamples and in
all different versions.9 This led them to propose a more complex
specificationinvolving interaction between A(t), H, g, and also the
gap between the individual country's level of A(t) and that of the
leading country. Such a specification opens up multiple channels
for the human capital variable to have impact on growth, which
then allows the theoretical propertiesof the human capitalvariable
to be better reflected in the regression results. While the issue of
inadequacy of the currently available variables as measures of
human capital across countries cannot be belittled, the approach
taken by Benhabib and Spiegel and others is certainly more
promising. The analysis in the next section indeed shows that
human capital is closely related to the estimated values of the A(O)
term. Benhabib and Spiegel, however, limit their analysis to single
cross-section regression with some variables entering in the first
differencedform. We intend to extend this approachto the panel
framework in a future paper. Empirical work has so far clearly
established that human capital plays a very important role in the
growth process. However, the question that remains still unresolved is, In WhatExact Way?
In sum, therefore, the above exercise shows that, despite the
pitfalls encountered regardingthe role of human capital, the effect
of controlling for the differences in the A(O)term remains robust.
The main properties of the panel results remain unchanged
whether or not we include human capital in the regression.
VIII.
ESTIMATED COUNTRY EFFECTS: A TENTATIVE ANALYSIS
Panel estimation permits us not only to allow for the individual country effects in the estimation of the other parameters of
the model, but also to get the estimates of these effects themselves.
In the case of LSDV estimation the recovery of the estimated
9. As they reported, ". . . the coefficient for human capital is insignificant and
enters with the wrong sign. . . . this result is independent of whether we use the
Kyriacou, Barro-Lee, or literacy data sets as proxies for the stock of human capital
in computing the growth rates of human capital" [Benhabib and Spiegel 1994,
p. 154].

1155
individual effects is direct because they are the estimated coefficients of the country dummies. If LSDV is implemented in the form
of "within-regression"-as
we do-the estimated country effects
are obtained as
(19)
pi =i
P3Xi-
9,
where
l
i=
TKiY
1 T-1
Yi,-1 =
X =
T
I
with t being the estimates of the time effects. In the case of the
correlated effects model, the MD estimation procedure yields
estimates of Kt. These can now be substituted back into equation
(14) to get the estimated country effects.
The estimated values of ,ui's are presented in column 3 of
Appendix 1.10The implied values of In A(O)i can be recovered from
the 'ii's using the formula, ki = (1 - e-XT) lnA(O)i. These are
presented in column 4. The dispersion and relative position of the
countries can be highlighted by computing the values of A(O)i and
expressing them relative to A(O)min(in the present case, the A(0) of
Somalia). These ratios may be called the A(0) index, and they are
presented in column 5. The rank of the countries in terms of this
index can be seen in column 6. The value of the index ranges from 1
to about 40, showing that the countries do vary enormously in
terms of their A(O) values. In general, however, we find a bottomheavy distribution. If we classify the countries according to whether
the estimated A(0) index is less than 5, between 5 and 10, 10 and
15, 15 and 20, 20 and 25, and greater than 25, and name the
corresponding groups for easy reference as I (Very Low), II (Low),
III (Medium), IV (High), V (Very High), and VI (Super High), then
we find that 63 (i.e., 66 percent) of the countries of the sample fall
into the lowest two groups. The histogram in Appendix 2 gives a
more visual display of the distribution along with Appendix other
information that makes it easy to read off from it both the rank and
the value of the A(0) index for particular countries.
The A(0) index is, obviously, a measure of efficiency with
which the countries are transforming their capital and labor
resources into output and hence is very close to the conventional
10. The estimates presented in Appendix 1 are based on NONOIL results. The
relative rankings of the countries do not change when these country effects are
estimated on the basis of the INTER sample. Also, the MD and LSDV yield similar
estimates when rounded up to two decimal points.

1156
concept of total factor productivity (TFP). The important difference is that while the TFPs are computed for the individual
countries on the basis of their respective time series data, the
country effects are inherently based on cross-country comparison
and are not subsequent upon individualcountry-analysis.
Appendix 3 gives a two-way distribution of the countries in
terms of the A(O)index and continents. In general, it conforms to
what may be called the expected pattern. A large proportion (84
percent) of the Africancountries fall in the Very Low group. On the
other hand, most of the European countries are in the High and
Very High groups. Latin and CentralAmericancountries generally
tend to avoid the Very Low group, and instead belong mostly
(about half) to Group II. The Asian countries seem to be more
widely distributed across the groups. While a good number of them
(about one-third) fall into the Very Low group, many of the rest
make their way to the higher groups.
In recent years there has been renewed interest in in-depth
analysis of growth performance of individual countries (see, for
example, Young [1992, 1995]). In light of this interest, it may be
worthwhile to note certain "unexpected"or interesting aspects of
the results regardingthe A(O)index. First is the super high value of
this index for Hong Kong. Not only did it have the highest value,
but it surpassed by this measure the next ranking country
(Canada)by a wide margin. Hong Kong's super value of the A(O)
index, comparedwith that of Singapore, lends support to some of
Young's [1992] conclusions from his detailed analysis of the
comparativegrowth performanceof these two city states. Second,
the high values of the A(O)index for the United States and Canada
may to some extent dispel the thesis of productivity slowdown in
these countries, in particular, the United States. This is further
supported by the relatively low values of the A(O)index for both
Japan and Germany. Third, some of the countries did remarkably
well in terms of the A(O) index. Among these are, for example,
Israel, Trinidad and Tobago, and Venezuela and also to some
extent, Syria and South Africa.1"Fourth, equally noteworthy is the
poorA(O)index of some other countries. Among the latter are, for
example, Korea, Chile, Portugal, and India. This result, particularly regardingKoreaand Chile, may draw some attention.
Having seen the range, dispersion, and ranking, we may next
11. However, it may be necessary to check into the possible role of extraction of
oil and other mineral deposits in this regard, particularly with respect to countries
like Venezuela and South Africa.

1157

5
1y
- 7
ocb
0o
0 00
0
0
~~~~~~~~~
0000
000~~04J0
ln0
6"
%00
00~~~~~~~~~~~~~-1
OD0
I
I_
1ny8
ODOI&0
l~8
oec OP*
0 0
O~~~~~~~AOL
? ?
0c00
?
7IW94
FIGURE I
Scatter Plot Matrix of in y6O,in aO, and in y85
try to look at the correlates, determinants, and implications of the

estimated A(O)'s. Here again, the approach can be similar to that
applied to the conventional TFPs. First of all, we note that A(O)i is
part of the production function, and hence it should be correlated
Also note that since

with the income levels of the countries.
A(O)4
is
the time-invariant term, it should be correlated with the income

levels of all the time periods considered in the sample. To check
this, we try to see how the estimated A(0)i values are correlated
with the per capita income levels of the beginning and end points of
the sample period, i.e., of 1960 and 1985. According to the
production function (equation (1)), we should have
(20)
y(t)
((~g)1-k(~
where, as before, y(t) is per capita income at time t and k(t) is per
capita capital stock. The scatter plot matrix presented in Figure I
shows the relationship between the log of A(O)(denoted by ln aG)
and the log of y of 1960 and 1985 (denoted in the figure by lniy6
and lny85, respectively). It shows quite clearly that ln aO is very

1158
closely associated with both in y6O and in y85. The simple correlation coefficients are 0.8656 and 0.9599 for they's of 1960 and 1985,
respectively. Note that the association between ln y60 and ln y85
themselves is very strong. This can be seen from the upper right or
lower left blocks of the scatter plot matrix. The simple correlation
coefficient between the two is 0.9202.12
Of course, one weakness of the above evidence is that we do not
have independent measures of A(0). The values of A(0) used to
determine the correlations are the outcome of an estimation
process in which values of y(t) themselves (along with other
variables) served as the data. Therefore, there may be an induced
element in the correlations cited above. However, even after
discounting for such a possibility, the strength of the correlations
probably remains very strong.
Next we turn to the issue of growth rates. Is the A(0) term
important in explaining growth? According to the Solow-CassKoopmans model, steady state growth is given by the exogenous
rate of technical progress. Hence, the focus here is on growth in
transition. Can A(0) affect transitional growth? Note that since
A(O) enters the production function in a multiplicative way, upon
log differencing of y(t) of any two time periods it would vanish. As
we can see from equation (20),
(21)
lny(t2) - lny(t2)
(1
0042
t1) +
a(ln k(t2) - In k(tj)).
There is no A(0) term on the right-hand side of the above equation.

This may create the impression that A(0) should not have any
effect on growth. However, according to the model, the countries
move toward their respective steady states, and, once that is taken
into consideration, the situation changes. We have already seen in
equation (10) that A(0) appears as one of the right-hand-side
variables in explaining the dynamics around steady state. However, in that equation, y(t1) also appears as another explanatory
variable, and sincey(t1) carries anA(0) term embodied in it, the role
of A(0) in the relationship gets somewhat confounded. We can
make this relationship clearer in the following way. The formula
12. This strong correlation between in y60 and in y85 is noteworthy. It seems
to suggest that the ranking of an economy in terms of per capita income in 1960 is
almost sufficient to predict its corresponding rank in 1985. These correlations also
suggest that persistent difference in the steady state levels of income because of
differences in A(0) is the most salient feature of cross-country growth. The
differences in A(0) outweigh the impact of relative changes across countries in other
variables, like saving rate etc. Put bluntly, improvement in A(O) seems to be more
important than raising the saving rate in changing the relative position of a country.

1159
ln-aO
SGP
HKG
JPN
KOR
.05 -
BWA
OMRA
THA
PPT
ESP0
COG EGY
SYPINI
TP*AR
ChLPPYIRL
KCAN
J
SUP
MUSDOM LI
USA
MWI
.
XAUS
GSP
Z
E
LAMIO
L
YE
P
ZE
E
PS
G
NG N$p
SWODB
On
JW~~~~~~~~~~MPPK
Go
CD
CO0
CD
TZA
~~~~~MPT
CD
LSP
URY
FIGURL
mo~G
ZAP GHA
TCD
-.05
Plot
Scate
GflO
vesuEN68
FIGURE II
Scatter Plot of In aO versus GR6085
for the steady state level of income is given by equation (4). In

multiplicative form this is equivalent to
(22)
A()e
Yt~~
+ +
n +g + 6
OA/(-a)
If we divide this by the expression for y(to) as in equation (20) and

set to equal to 0, then we get
(23)
Yo = A(0)aegt
(s/(n + g +))a/(
gk
a)
The other two multiplicative terms on the right-hand side of the

above expression are positive, which shows clearly that, other
things being equal, the distance between the steady state and
initial level of income increases with the increase in A(0). However,
we also know that, other things being equal, the more distant a
country is from its steady state, the higher the growth rate. Hence,
there should be a positive relationship between the observed
growth rates and the estimated values of A(O). Figure II gives a
scatter plot of the growth rate of per capita GDP between 1985 and
1960 (GR6085) and In aG. We find a strong positive relationship
between the two (the simple correlation coefficient is 0.6486). The

1160
line in the plot shows the fitted values from the bivariate regression
of GR6085 on In a0.13 This plot also helps to illustrate the relative
position of the particular countries in the In a0 scale and the
corresponding growth performance. (The codes to the abbreviated
country names used in the plot can be seen in Appendix 1.)
The above analysis therefore shows that higher values of A(0)
are associated not only with higher levels of per capita income, but
also with higher growth rates. The important question then is
what the determinants of A(0) are. In our discussion in Section II,
we have considered the qualitative list of the possible components
of A(0). The question, however, is whether we can determine (at
least approximately) their relative quantitative significance. This is
obviously a big question, and we cannot adequately address it
within the limits of the current paper. However, it indicates one of
the lines along which future research can proceed.14
Before concluding the discussion of this section, however, we
want to look into the relationship of A(0) with one other variable of
particular interest, namely, human capital. This has added relevance in view of our discussion in Section VII. We use two different
measures of human capital for this purpose. One is the SCHOOL
variable used by M-R-W and the other is Barro and Lee's HUMAN
variable. We have alluded to both these variables in our discussion
in the previous section. Recall that HUMAN in Barro and Lee is a
panel variable. For our current cross-sectional analysis, we compute averages from the data on HUMAN for different quinquennial
years of the respective countries. These can be seen in the last
column of Appendix 1. Figure III gives the scatter plot matrix of
ln a, SCHOOL, and HUMAN. The plot shows that both the
SCHOOL and HUMAN variables are indeed very closely related to
In aG. The positive association is closer to HUMAN than to
SCHOOL. Simple correlation coefficients of In a0 with SCHOOL
and HUMAN are 0.7120 and 0.7802, respectively. (The correlation
coefficient between SCHOOL and HUMAN is 0.7529.)
This correlation between In a0 with human capital variables
13. The regression has the following results:
GR6085
-7.5296
(1.1429)
+ 1.3166 lnaO;
(0.1594)
N = 96,
0.4145
F194 = 68.26.
The figures in parentheses are standard errors. This indicates that an increase in
In aO by 1, which in turn implies about a threefold increase in the value of A(0),
would be associated with a 1.32 percentage point increase in the growth rate.
14. The author is currently engaged in such research and intends to produce a
paper on this topic in future.

1161

5
9
15
school
0
scho0
0
a%
05
~~~~~~~0
0 0
~~~~~~
00
0 0
0510
0O~~~%
00
o~~~oR
o~~~~~~o~
~
8
0
9
~~
0~~~0
oo~~~,,o
10
-10
8 ,000070000
~~~ of SCOL
Oc~~~Q3~ Scatter~
0
0 Plo Mari
0
l0 a,
an5UA
0~~~~~~~~~~~~o
obtain
e
i
10
0
15
-0~0
ha
~~~~~~0
-5
51
(single) cross-section regression, the variation in the SCHOOL

variable interacts with only the cross-sectional variations of other
variables. In a panel framework, on the other hand, the individual
effects force the regression results to be derived more from the
within variation (which is what conditional convergence is supposed to capture). This produces the higher rates of convergence,
something that was not obtained in M-R-W.
This also helps illuminate the question that we posed toward
the end of Section IV, namely, "In What Exact Way" does human
capital play its role in the growth process? The close relation
between the estimated values of ln A(O)i and either SCHOOL or
HUMAN variable seems to suggest that the channel through which
human capital affects growth may be more tortuous than that
implied by its simple multiplicative inclusion in the production
function with a different exponent. Instead, evidence indicates that

1162
this channel most likely runs through A. Of course, this does not
resolve the question quoted above, but it perhaps at least indicates
where to look for the answer.
IX. CONCLUSION
We argued for and implemented a panel data approach to the
study of cross-country growth, in particular, of the phenomenon of
convergence. This results in higher rates of conditional convergence and lower values of the elasticity of output with respect to
capital. These results are explainable in terms of correction for
omitted variable bias involved with the single cross-section regression. From growth theory's point of view, they highlight the
significance of the differences in the aggregate production functions.
However, the faster convergence that we observe is conditional. So to speak, this reveals the flip side of our finding. By being
more successful (through the panel framework) in controlling for
further sources of difference in the steady state levels of income, we
have, at the same time, made the obtained convergence hollower.
This is because convergence is more commonly understood as
different countries of the world approaching the same or similar
levels of income (i.e., in the "absolute" sense). There is probably
little solace to be derived from finding that countries in the world
are converging at a faster rate, when the points to which they are
converging remain very different.
This also sheds light on the issue of policy activism. The faster
rate of convergence may appear to reinforce the policy-irrelevance
ethos ascribed to the Solow model. In actuality, however, the
opposite is the case. Traditionally, only the saving and population
growth rates were thought to be the variables for policies to be
directed to. However, our study highlights the role of the A(O) term
as a determinant of the steady state level of income. It thus brings
to the fore the fact that, even with similar rates of saving and
population growth, a country can directly improve its long-run
economic position by bringing about improvements in the components of A(0). Also, improvements in A(O) can have salutary effects
on s and n leading to a further (indirect) increase in the steady state
level of income. Therefore, this study points to a richer scope for
policies in raising the long-run income levels of countries and in
quickening the pace of reaching them. One of the distinguishing
aspects of development economics as a branch of economics has
been its emphasis on and discussion of the elements that enter into

1163
GROWTH EMPIRICS: A PANEL DATA APPROACH
A(O). Consequently, the current study helps to connect that

discussion with the recent work on growth empirics.
Finally, the individual country effects estimated through the
panel procedure are another kind of comparative measure of total
factor productivity (TFP), or the efficiency with which other factors
of production are being converted into output. Just as the conventional TFPs computed for individual countries from their respective time series data can form the material for subsequent crosscountry analysis, the estimated individual country effects also
provide a new point of departure for a similar kind of analysis. In
particular, we may be interested in ascertaining the relative
strength of correlation of different determinants of the A(O) term
that we have discussed above.
APPENDIX 1: ESTIMATEDCOUNTRYEFFECTS AND HUMAN CAPITALFOR

DIFFERENT COUNTRIES
A(0)j
Country
Code
Algeria
Angola
Benin
Botswana
Burundi
Cameroon
Ctrl.Afr.Rep.
Chad
Congo
Egypt
Ethiopia
Ghana
Ivory Coast
Kenya
Liberia
Madagascar
Malawi
Mali
Mauritania
Mauritius
Morocco
Mozambique
Niger
Nigeria
Rwanda
Senegal
Sierra Leone
DZA
AGO
BEN
BWA
BDI
CMR
CAR
TCD
COG
EGY
ETH
GHA
CIV
KEN
LBR
MDG
MWI
MLI
MRT
MUS
MAR
MOZ
NER
NGA
RWA
SEN
SLE
i
1.45
1.38
1.25
1.47
1.23
1.42
1.20
1.14
1.36
1.41
1.27
1.19
1.43
1.25
1.21
1.29
1.21
1.20
1.16
1.45
1.56
1.36
1.27
1.30
1.23
1.34
1.26
In A(0)i
A(0)min
Rank
SCHOOL
HUMAN
6.97
6.63
6.00
7.06
5.91
6.82
5.76
5.48
6.53
6.77
6.10
5.72
6.87
6.00
5.81
6.20
5.81
5.76
5.57
6.97
7.49
6.53
6.10
6.24
5.91
6.44
6.05
5.12
3.49
1.96
5.64
1.78
4.43
1.54
1.16
3.32
4.23
2.16
1.47
4.65
1.96
1.62
2.37
1.62
1.54
1.27
5.12
8.68
3.32
2.16
2.49
1.78
3.17
2.06
53
65
81
51
83
60
88
94
68
62
76
90
58
80
87
74
86
89
91
54
37
67
77
73
84
69
78
4.5
1.8
1.8
2.9
0.4
3.4
1.4
0.4
3.8
7.0
1.1
4.7
2.3
2.4
2.5
2.6
0.6
1.0
1.0
7.3
3.6
0.7
0.5
2.3
0.4
1.7
1.7
1.27

0.57*
1.96
1.75
2.71*
1.32*
1.93
1.76
1.04
2.04
0.47*
3.55
0.71
0.61
0.86*
1.81
1.17
1164
APPENDIX
(CONTINUED)
Country
Somalia
S. Africa
Sudan
Tanzania
Togo
Tunisia
Uganda
Zaire
Zambia
Zimbabwe
Bangladesh
Burma
Hong Kong
India
Israel
Japan
Jordan
S. Korea
Malaysia
Nepal
Pakistan
Philippines
Singapore
Sri Lanka
Syria
Thailand
Austria
Belgium
Denmark
Finland
France
Germany
Greece
Ireland
Italy
Holland
Norway
Portugal
Spain
Sweden
Switzerland
Turkey
U. Kingdom
Canada
Code
pi
In A(0)j
A(0)i
A(0)min
Rank
SCHOOL
HUMAN
SOM
ZAF
SDN
TZA
TGO
TUN
UGA
ZAR
ZMB
ZWE
BGD
BUR
HKG
IND
ISR
JPN
JOR
KOR
MYS
NPL
PAK
PHL
SGP
LKA
SYR
THA
AUT
BEL
DNK
FIN
FRA
DEU
GRC
IRL
ITA
NLD
NOR
PRT
ESP
SWE
CHE
TUR
GBR
CAN
1.11
1.60
1.22
1.15
1.25
1.53
1.33
1.15
1.14
1.33
1.38
1.29
1.89
1.25
1.70
1.75
1.52
1.60
1.60
1.36
1.46
1.45
1.77
1.41
1.64
1.51
1.72
1.75
1.74
1.66
1.75
1.72
1.60
1.60
1.69
1.73
1.77
1.58
1.75
1.73
1.70
1.53
1.73
1.81
5.33
7.69
5.86
5.52
6.00
7.35
6.39
5.52
5.48
6.39
6.63
6.20
9.08
6.00
8.17
8.41
7.30
7.69
7.69
6.53
7.01
6.97
8.50
6.77
7.88
7.25
8.26
8.41
8.36
7.97
8.41
8.26
7.69
7.69
8.12
8.31
8.50
7.59
8.41
8.31
8.17
7.35
8.31
8.69
1.00
10.52
1.70
1.21
1.96
7.89
2.88
1.21
1.16
2.88
3.66
2.37
38.49
1.96
17.01
21.63
7.17
10.03
10.52
3.32
5.37
5.12
23.81
4.23
12.75
6.83
18.73
21.63
20.61
14.04
21.63
18.73
10.52
10.52
16.21
19.65
23.81
9.56
21.63
19.65
17.09
7.52
23.82
28.85
96
28
85
93
82
41
72
92
95
71
63
75
1
74
17
10
44
33
31
66
52
55
6
61
25
47
15
8
11
23
9
14
30
21
19
12
5
34
7
13
18
43
4
2
1.1
3.0
2.0
0.5
2.9
4.3
1.1
3.6
2.4
4.4
3.2
3.5
7.2
5.1
9.5
10.9
10.8
10.2
7.3
2.3
3.0
10.6
9.0
8.3
8.8
4.4
8.0
9.3
10.7
11.5
8.9
8.4
7.9
11.4
7.1
10.7
10.0
5.8
8.0
7.9
4.8
5.5
8.9
10.6
4.46
0.50
2.02
0.96
1.28
1.36
1.32
2.70
2.08
1.21
1.27
5.85
2.19
8.04
7.39
2.46
5.65
3.74
0.33
1.46
5.12
3.72
4.68
2.24
3.81
5.42
8.16
9.82
8.58
5.21
8.15
5.63
6.96
5.34
7.23
8.19
2.34
4.53
8.25
7.49
2.35
7.92
9.16


APPENDIX
1165
(CONTINUED)
Country
Code
pi
In A(0)i
A(0)i
A(0)min
Rank
SCHOOL
HUMAN
Costa Rica
Dom. Rep.
El Salvador
Guatemala
Haiti
Honduras
Jamaica
Mexico
Nicaragua
Panama
Trinidad
United States
Argentina
Bolivia
Brazil
Chile
Colombia
Ecuador
Paraguay
Peru
Uruguay
Venezuela
Australia
New Zealand
P. N. Guinea
CRI
DOM
SLV
GTM
HTL
HND
JAM
MEX
NIC
PAN
TTO
USA
ARG
BOL
BRA
CHL
COL
ECU
PRY
PER
URY
VEN
AUS
NZL
PNG
1.60
1.48
1.51
1.56
1.35
1.37
1.43
1.65
1.55
1.54
1.70
1.80
1.52
1.43
1.62
1.49
1.54
1.50
1.57
1.57
1.60
1.67
1.69
1.69
1.44
7.69
7.11
7.25
7.49
6.48
6.58
6.87
7.93
7.45
7.40
8.17
8.65
7.30
6.87
7.78
7.16
7.40
7.21
7.54
7.54
7.69
8.02
8.12
8.12
6.92
10.52
6.20
6.83
8.68
3.17
3.49
4.65
13.38
8.28
7.89
17.01
27.50
7.17
4.65
11.58
6.20
6.51
9.11
9.11
10.52
14.73
16.21
16.21
10.52
4.88
27
50
46
38
70
64
59
24
39
42
16
3
45
57
26
49
40
48
36
35
32
22
20
21
56
7.0
5.8
3.9
2.4
1.9
3.7
11.2
6.6
5.8
11.6
8.8
11.9
5.0
4.9
4.7
7.7
6.1
7.2
4.4
8.0
7.0
7.0
9.8
11.9
1.5
4.26
3.17
2.56
1.85
1.12
2.30
3.25
3.32
2.59
5.09
5.36
10.44
5.92
3.75
2.90
5.57
3.59
4.04
4.04
4.27
5.26
3.69
9.72
10.70
1.13
A period (.) indicates a missing value of the HUMAN variable, and an asterisk (*) indicates that the full set
of quinquennialvalues of the HUMANvariablewere not availablefor the countryand the averagehas been
computedon the basis of fewerdatapoints.

1166
APPENDIX
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
PNG
BOL
CIV
JAM
CMR
LKA
EGY
BGD
HND
AGO
NPL
MOZ
COG
SEN
HTI
ZWE
UGA
NGA
MDG
BUR
ETH
NER
SLE
IND
KEN
BEN
TGO
BDI
RWA
SDN
MWI
LBR
CAF
MLI
GHA
MRT
ZAR
TZA
TCD
ZMB
SOM
(4.9)
(4.7)
(4.7)
(4.7)
(4.4)
(4.2)
(4.2)
(3.7)
(3.5)
(3.5)
(3.3)
(3.3)
(3.3)
(3.2)
(3.2)
(2.9)
(2.9)
(2.2)
(2.4)
(2.4)
(2.2)
(2.2)
(2.1)
(2.0)
(2.0)
(2.0)
(2.0)
(1.8)
(1.8)
(1.7)
(1.6)
(1.6)
(1.5)
(1.5)
(1.5)
(1.3)
(1.2)
(1.2)
(1.2)
(1.2)
(1.0)
1-5
PRT
PER
PRY
MAR
GTM
NIC
COL
TUN
PAN
TUR
JOR
ARG
SLV
THA
ECU
CHL
DOM
BWA
PAK
DZA
MUS
PHL
(9.6)
(9.1)
(9.1)
(8.9)
(8.7)
(8.3)
(7.9)
(7.9)
(7.9)
(7.5)
(7.2)
(7.2)
(6.8)
(6.8)
(6.5)
(6.2)
(6.2)
(5.6)
(5.4)
(5.1)
(5.1)
(5.1)
5-10
2:
HISTOGRAM OF THE
VEN
FIN
MEX
SYR
BRA
CRI
ZAF
IRL
GRC
MYS
URY
KOR
(14.7)
(14.0)
(13.4)
(12.8)
(11.6)
(10.5)
(10.5)
(10.5)
(10.5)
(10.5)
(10.5)
(10.0)
10-15
NLD
SWE
DEU
AUT
TTO
ISR
CHE
ITA
AUS
NZL
A(0) INDEX
(19.7)
(19.7)
(18.7)
(18.7)
(17.0)
(17.0)
(17.0)
(16.2)
(16.2)
(16.2)
15-20
GBR
NOR
SGP
ESP
BEL
FRA
JPN
DNK
(23.8)
(23.8)
(23.8)
(21.6)
(21.6)
(21.6) HKG (38.5)
(21.6) CAN (28.9)
(20.6) USA (27.5)
20-25
25+
The figures in parentheses are the corresponding values of the A() index.
The full names corresponding to the abbreviated names of the countries can be seen in Appendix 1.

1167

APPENDIX
3:
OF THE COUNTRIES IN TERMS OF THE
DISTRIBUTION
A(0)
INDEX AND
CONTINENT
Continent
Group
I
Group
II
Group
III
Group
IV
Group
V
Group
VI
Total
Africa
Americaa
Asiab
Europec
31
4
6
0
5
11
4
2
1
5
3
3
0
1
3
6
0
0
2
6
0
2
1
0
37
23
19
17
Total
41
22
12
10
96
a. America includes Latin, Central, and North American countries.

b. Asia includes Papua New Guinea, New Zealand, and Australia.
c. Turkey is included in Europe.
DISTANCE
ESTIMATION
APPENDIX4: NOTEON MINIMUM
The purpose of Appendix 4 is to give some details on the
Minimum Distance estimation method used in this paper. The
arguments for use of this method and its basic principles and
properties have been discussed in the text. Therefore, we concentrate here only on its operational aspects. Further details can be
seen in Chamberlain [1982, 1983].
The restricted model is
=it=
Yit-
+ Pxit + ki + Vito
where xi, is the notation for [in (s) - in (n + g + r)]. We ignore the
time effect term 9t which is taken care of by appropriate time
dummies.
This model is augmented by the following specifications for pi
andyio:
i=
+
K0Xij
Ko
YiO = ko +
+ K2Xi2 +
kXil + 4)2Xi2 +
+ KTXiT+ PIi
+ 4)TXiT + hi,
where E[4ijlxi, . . . XiT] = 0 and E[4 Ixi1. . . XiT]
0.
The MD estimator involves first substituting out the lagged

dependent variable by repeated substitution, so that equations for
all the periods are functions of the exogenous xit's only (apart from
pi and Yio). So in our case with T = 5, this process leads to the
following reduced-form equations (we suppress the cross-section
subscript i):
Y1=
AX1 + YY0+ RL+ Vi
Y2 = 'SYPX1+ r3X2 + Y2Yo+ (R + mYa)+ (v2 + _yV1)

1168
+ -YPX2+ 3X3+ Y3Y0+ (p + _yp + _y2p)
y3 =2pX
+ (V3 + YV2+ y2V,)

y4
y3X1 + _y2 X2 + _YPX3+ PX4 + Y4Y0

+ (pL+ yp. + -y2p + y3R) + (v4 + yv3 + y2V2 + y3v,)
Y5 = y43X + _y3X2 + y2r3X3+ _Yx4 + 3X5 + Y5Yo

+ (R + ,yi + ,y2p + ,y3p + y4R)
+ (v5 + yv4 + y2v3 + y3v2 + y4V1).
In matrix form this may be expressed as

Y1
Y2
Y3
Y4
Y5
Syj3
B
Y213
ly3
y41
0
3
0
0
?0
0
_y3
y2
13
_y
y3r
_y2r
x1
x2
x3
x4
y4
_y1
p
X5
y5
Y3 Y0
1
1+ y
U1
U2
+ Ut3
u4
U5
l+ y+ y2
1+y+y2+y3
1+y+y2+y3+y4
where ut stands for the respective composite error term consisting

of vi's. Now, if we substitute for yo and ,u per their specifications
adopted above, then we get the Il-matrix of the reduced-form
coefficients of the following form:
IP
(3 0
-y3
~2p _ p
'y$3 _y3
p~2 _p
0 0
0 0
~2
0 +
03
y
y
1+y
1 +l+ y++ y2 + y3
y y2 )K.y
+ y + y2 + y3 + y4
If xe's are strictly exogenous, then they are uncorrelated with the

1169
ut's, and the reduced-form equations can be estimated by applying

ordinary least squares (OLS).
As can be seen from above, if we ignore the intercept term,
then the Hl-matrix in the present case is a (5 x 5) matrix, so that
altogether it has 25 elements which are nonlinear functions of
twelve underlying coefficients, namely, (ypKlK2K3K4K5+l+2+3+4+5),
which we may denote by the vector 0'. The purpose of MD
estimation is to impose the restrictions and squeeze out an
estimate of 0 from that of H by minimizing
0 = argmin (vecH - g(0))'Aj1(vecHI - g(O)),

where g(O) is the vector valued function mapping the elements of 0
into vec(H). Chamberlain showed that the optimal choice for the
weighting matrix AJ-1is the inverse of
fl =
E[(y, - HIxi)(yi - H0xj)' 0
-l(x'x/
l'],
where HOis the matrix of true coefficients and Fx = E(x ixj). It may
be noted that ft is a heteroskedasticity-consistent weighting matrix. In actual implementation we need to use its consistent sample
analog,
Ni=
z[(Yi
- Ji'xj)(Yi-
ix)' 0 S;-(xixltS;1],
where Sx = V xixI'/N. Minimization of equation (21) is attained

through an iterative procedure, and we use the modified GaussNewton algorithm to implement it.
HARVARDUNIVERSITY
REFERENCES
Amemiya, Takeshi, "A Note on the Estimation of Balestra-Nerlove Models,"
Technical Report No. 4, Institute for Mathematical Studies in Social Sciences,
Stanford University, 1967.
Barro, Robert J., "Economic Growth in a Cross Section of Countries," NBER
Working Paper No. 3120, Cambridge, September, 1989.
Barro, Robert J., and Jong-Wha Lee, "International Comparisons of Educational
Attainment," Journal of Monetary Economics, XXXII (1993), 363-94.
Barro, Robert J., and Xavier Sala-i-Martin, "Convergence," Journal of Political
Economy, C (1992),223-51.
Baumol, William J., "Productivity Growth, Convergence and Welfare: What the
Long-Run Data Show?" American Economic Review, LXXVI (1986), 1072-85.
Benhabib Jess, and Mark M. Spiegel, "The Role of Human Capital in Economic
Development: Evidence from Aggregate Cross-Country Data," Journal of
Monetary Economics, XXXIV (1994), 143-73.

1170
Cass, David, "Optimum Growth in an Aggregative Model of Capital Accumulation,"

Review of Economic Studies, XXXII (1965), 233-40.
Chamberlain, Gary, "Multivariate Regression Models for Panel Data," Journal of
Econometrics, XVIII (1982), 5-46.
"Panel Data," in Handbook of Econometrics, Zvi Griliches and Michael
Intriligator, eds. (Amsterdam: North-Holland, 1983), pp. 1247-1318.
Chua, Hak B., "Regional Spillovers and Economic Growth," Department of
Economics, Harvard University, 1992.
De Gregorio, Jose, "Economic Growth in Latin America," Journal of Development
Economics, XXXIX (1992), 59-84.
De Long, J. Bradford, "Productivity Growth, Convergence, and Welfare, American
Economic Review, LXXIX (1989), 1010-30.
Durlauf, S., and P. A. Johnson, "Local and Global Convergence across National
Economies," Department of Economics, Stanford University, 1991.
Islam, Nazrul, "Estimation of Dynamic Models from Panel Data: A Review,"
Department of Economics, Harvard University, 1991.
,"Small Sample Performance of Dynamic Panel Data Estimators: A Monte
Carlo Study," Department of Economics, Harvard University, 1992.
King, Robert G., and Sergio T. Rebelo, "Transitional Dynamics and Economic
Growth in the Neoclassical Model," NBER Working Paper No. 3185, Cambridge, 1989.
Koopmans, T. C., "On the Concept of Optimal Economic Growth," in The Economic
Approach to Development Planning, Pontifical Academy of Sciences (Amsterdam: North-Holland, 1965).
Maddala, G. S., "The Use of Variance Components Model in Pooling Cross-Section
and Time Series Data," Econometrica, XXXIX (1971), 341-58.
Maddison, A., Phases of Capitalist Development (Oxford: Oxford University Press,
1982).
Mankiw, N. Gregory, David Romer, and David Weil, "A Contribution to the
Empirics of Economic Growth," Quarterly Journal of Economics, CVII (1992),
407-37.
Mundlak, Yari, "On the Pooling of Time Series and Cross-Section Data," Econometrica, XXXVI (1978), 69-85.
Rebelo, Sergio, "Long-Run Policy Analysis and Long-Run Growth," Journal of
Political Economy, XCIX (1991), 500-21.
Romer, Paul, "Capital Accumulation in the Theory of Long-Run Growth," in
Modern Business Cycle Theory, Robert J. Barro, ed. (Cambridge, MA: Harvard
University Press, 1989a).
,"Human Capital and Growth: Theory and Evidence," NBER Working Paper
No. 3173, Cambridge, 1989b.
Solow, Robert M., "A Contribution to the Theory of Economic Growth," Quarterly
Journal of Economics, LXX (1956), 65-94.
Summers, Robert, and Alan Heston, "A New Set of International Comparisons of
Real Product and Price Levels Estimates for 130 Countries, 1950-85," Review
of Income and Wealth, XXXIV (1988), 1-26.
Young, Alwyn, "A Tale of Two Cities: Factor Accumulation and Technical Change
in Hong Kong and Singapore," in NBER Macroeconomics Annual 1992, Olivier
J. Blanchard and Stanley Fischer, eds. (Cambridge, MIT Press, 1992), pp. 1354.
"The Tyranny of Numbers: Confronting the Statistical Realities of the East
Asian Growth Experience," Quarterly Journal of Economics, CX (1995),
641-80.


Islam (1995) - Growth Empirics A Panel Data Approach

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Islam (1995) - Growth Empirics A Panel Data Approach

Uploaded by

Copyright:

Available Formats

Growth Empirics: A Panel Data Approach

Author(s): Nazrul Islam

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATA APPROACH*

In recent years there has been considerable empirical work on

*I would like to thank Dale W. Jorgenson, Gary Chamberlain, Guido W.

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

cally difficult to allow for such differences in the production

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

(country) effects. Improvements in these factors may have direct

II. THE ISSUE OF "CONVERGENCE"AND ITS EMPIRICALSEARCH

A major focus of recent work on growth empirics has been the

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

and Rebelo 1989]. The rise of endogenous growth theories, as is

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

The tests for convergence conducted so far have followed a

The usefulness of a panel data approachcan be illustrated on

where Y is output, K is capital, and L is labor. L and A are assumed

to grow exogenously at rates n and g so that

Assuming that s is the constant fraction of output that is saved and

k(t) = s9(t) - (n + g + 6)k(t)

where 8 is the constant rate of depreciation. It is evident that k

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

Upon substitution this gives the following expression for steady

Assuming that the countries are currently in their steady states,

where a is a constant and e is the country-specificshift or shock

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

QUARTERLY JOURNAL OF ECONOMICS

into the constant term a, they derived the specification:

At this stage, however, M-R-W made the assumption that E is

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

+ g + 8)(1 - oa).This equation implies that

In9* + e-AT In9(t1),

1n9(t2) - In9(t1) = (1 - e-At) In 9* - (1 -

In the standard partial adjustment model, the "optimal" or

1n9(t2) - In 9(t1) = (1 - e -AT)

M-R-W used this equation to study the process of convergence

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

income per capita. We may, therefore, reformulate the equation in

Iny(t2) - Iny(t1) = (1 - e-T)

However, if we collect terms with In y(t1) on the right-hand side, we

It can now be seen that the above represents a dynamic panel

Yit = YYit-i + E jijxI't+ nt + pi + vit,

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

such estimation is whether the individualeffects are to be thought

oo. However, the

asymptoticpropertiesof panel data estimators can be consideredin

One of the reasons for the recent surge in work on growth

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

data set. In fact, the emergence of new (endogenous) growth

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

1 is 1960, and saving and population growth variables are averages

This content downloaded from 143.107.92.181 on Fri, 25 Apr 2014 17:43:38 PM

GROWTHEMPIRICS:A PANEL DATAAPPROACH

Wald test of restriction: